Start with a repeated process
Look for work with a clear trigger, recognisable inputs and an observable outcome. Describe what happens today before choosing a tool.
Separate the steps
Some steps need language interpretation; others only need a rule or a connection between systems. Record where people exercise judgement and why.
Establish a baseline
Measure volume, handling time and exceptions over a representative period. Time released is not automatically a cash saving: consider how that capacity will be used.
Choose a bounded pilot
Select a scope that can be evaluated without changing every system at once. Agree what success and an unacceptable result look like before implementation.